SenseNova-6.8-Code2Video — GGUF

GGUF quantizations of sensenova/SenseNova-6.8-Code2Video (27B hybrid qwen35 architecture), complete with the F16 Multimodal Vision Projector (mmproj-F16.gguf) and a grafted 1-layer Multi-Token Prediction (blk.64 / NextN) speculative decoding head preserved at Q8_0.

---

1. Embedded Q8_0 Multi-Token Prediction (MTP) Head (blk.64.*):
2. Hybrid Gated DeltaNet (ssm_*) & Attention Gate (attn_gate) Protection:
3. 100% Unique Multi-Domain imatrix Calibration:
   • Calibrated over 20 chunks (n_ctx=512) of 100% unique, non-repeating multi-domain data covering:
     • Bilingual (EN/ZH) Claude Code agent tool-calling trajectories & multimodal keyframe critiques
     • Remotion / React TypeScript multi-scene kinetic typography & data visualization code
     • Python procedural audio synthesis (numpy / scipy FM bells, Karplus-Strong pizzicato, Schroeder stereo reverb, FFmpeg filtergraphs)
     • Manim 3D differential geometry & Three.js custom GLSL shaders
4. Full Multimodal Vision Support (mmproj-F16.gguf):

---

📦 Provided Files, Sizes & Verified Perplexity (PPL)

💡 Which file should I pick?

• 16 GB VRAM GPUs (RTX 4080 / 4070 Ti Super / 5070 Ti): Grab IQ3_M-MTP.gguf (15.1 GB) if running alongside mmproj-F16.gguf, or IQ4_XS-MTP.gguf (17.2 GB) for text/code generation with partial offload.
• 24 GB VRAM GPUs (RTX 3090 / 4090 / 5090): Grab Q4_K_M-MTP.gguf (18.5 GB) or Q5_K_M-MTP.gguf (20.6 GB) + mmproj-F16.gguf (928 MB) for full GPU offload with long context.
• 32 GB+ Unified Memory / Workstation GPUs: Grab Q6_K-MTP.gguf (22.9 GB) or Q8_0-MTP.gguf (29.0 GB).

---

🚀 Quickstart & Usage (llama.cpp)

Make sure your llama.cpp build is up to date (supports the qwen35 architecture and llama-mtmd-cli).

1. Fast Download with huggingface-cli

# Download the recommended Q4_K_M quant + Vision Projector
hf download <your-username>/SenseNova-6.8-Code2Video-GGUF \
  SenseNova-6.8-Code2Video-Q4_K_M-MTP.gguf \
  SenseNova-6.8-Code2Video-mmproj-F16.gguf \
  --local-dir ./models

2. Text & Code2Video Generation (llama-cli)

./llama.cpp/build/bin/llama-cli \
  -m ./models/SenseNova-6.8-Code2Video-Q4_K_M-MTP.gguf \
  -ngl 99 \
  -c 32768 \
  --temp 0.6 \
  --top-k 20 \
  --top-p 0.95 \
  -p "Write a complete Python script using numpy and scipy to synthesize a stereo Karplus-Strong plucked string chord progression and save it to chord.wav."

3. Multimodal Vision & Keyframe Critique (llama-mtmd-cli)

./llama.cpp/build/bin/llama-mtmd-cli \
  -m ./models/SenseNova-6.8-Code2Video-Q4_K_M-MTP.gguf \
  --mmproj ./models/SenseNova-6.8-Code2Video-mmproj-F16.gguf \
  -ngl 99 \
  --image ./rendered_keyframe_0120.png \
  -p "Inspect this rendered animation keyframe for typography overlap, contrast issues, or layout clipping, and provide the exact CSS/Canvas fix."

4. OpenAI-Compatible Multimodal Server (llama-server for Claude Code / Agent Harnesses)

./llama.cpp/build/bin/llama-server \
  -m ./models/SenseNova-6.8-Code2Video-Q4_K_M-MTP.gguf \
  --mmproj ./models/SenseNova-6.8-Code2Video-mmproj-F16.gguf \
  -ngl 99 \
  -c 65536 \
  --port 8080 \
  --jinja

---

🔬 Architecture & Metadata Summary

• Architecture: qwen35 (Qwen3_5ForConditionalGeneration)
• Total Blocks: 65 (64 hybrid backbone layers + 1 grafted blk.64 NextN/MTP speculative prediction layer)
• Recurrent vs. Full Attention Pattern: Every 4th layer (blk.3, 7, 11, ..., 63) is Full Grouped-Query Attention (24 Q heads / 4 KV heads); the remaining 48 layers are Gated DeltaNet recurrent state-space layers (ssm.inner_size = 6144, ssm.state_size = 128, ssm.group_count = 16, ssm.time_step_rank = 48).
• Context Length: 262,144 tokens (YaRN RoPE scaling, rope.freq_base = 10,000,000)
• Vocabulary Size: 248,320 tokens

---

📜 License & Credits

• Original Model: sensenova/SenseNova-6.8-Code2Video by SenseTime Research (apache-2.0).
• MTP Donor Weights: Qwen/Qwen3.8-27B (apache-2.0).